基于客户细分和AdaBoost的电子商务客户流失预测研究  

E-commerce Customer Churn Prediction based on Customer Segmentation and AdaBoost

作　　者：武小军[1] 孟苏芳 WU Xiaojun MENG Sufang(School of Economics and Management, Tongji University, Shanghai 200092, Chin)

机构地区：[1]同济大学经济与管理学院,上海200092

出　　处：《工业工程》2017年第2期99-107,共9页Industrial Engineering Journal

摘　　要：为了准确识别高价值电子商务客户,提高对非流失客户的预测精度,本文首先对电子商务客户进行Kmediods聚类细分识别出高价值客户,再应用过采样和欠采样相结合的改进SMOTE处理不平衡的电子商务客户数据,最后用Ada Boost算法进行预测。实证研究表明,与成熟的客户流失预测算法BP神经网络、支持向量机(SVM)和改进支持向量机(CW-SVM)相比,该方法能更好地提高预测效果,与未细分前预测效果对比,客户细分后预测效果更好。In order to identify the high value customers as well as improve the prediction accuracy of non-churn customers, the high value customer groups are identified by K-mediods clustering and the churn data processed with improved SMOTE （synthetic minority oversampling teachnique）, which combines oversampling and undersampling methods to balance the datasets and generates the certain size of positive and negative samples by setting sampling ratio and controlling the model training time, then AdaBoost algorithm is employed to predict. At last, an empirical study on B2C E-commerce platform shows that the integrated model has better efficiency and higher prediction precision compared with the mature customer churn prediction algorithms, such as BP Neural, SVM （support vector machine） and CW-SVM （class weighted support vector machine）. Meanwhile, the prediction model of e-commerce customer churning based on customer segmentation has been proved to have better prediction performance.

关 键 词：客户细分 不平衡数据 SMOTE算法 ADA Boost算法 

分 类 号：F270[经济管理—企业管理]